diff --git a/lib/active_agent/providers/_base_provider.rb b/lib/active_agent/providers/_base_provider.rb
index 165fa8f7..f88bfd68 100644
--- a/lib/active_agent/providers/_base_provider.rb
+++ b/lib/active_agent/providers/_base_provider.rb
@@ -54,8 +54,11 @@ class ProvidersError < StandardError; end
       attr_internal :options, :context, :trace_id,   # Setup
                     :request, :message_stack,        # Runtime
                     :stream_broadcaster, :streaming, # Callback (Streams)
+                    :stream_completion_pending,      # Callback (Streams)
+                    :stream_completion_result,       # Callback (Streams)
                     :tools_function,                 # Callback (Tools)
                     :usage_stack,                    # Usage Tracking
+                    :stream_usage_index,             # Usage Tracking (Streams)
                     :max_tool_turns, :tool_turns     # Tool-loop safety
 
       # Upper bound on tool-calling round-trips within one generation. A
@@ -118,6 +121,9 @@ def initialize(kwargs = {})
         self.context            = kwargs
         self.message_stack      = []
         self.usage_stack        = []
+        self.stream_completion_pending = false
+        self.stream_completion_result  = nil
+        self.stream_usage_index        = nil
       end
 
       # Generates prompt preview without executing the API call.
@@ -179,6 +185,9 @@ def instrument(name, payload = {}, &block)
       #
       # @return [ActiveAgent::Providers::Common::PromptResponse]
       def resolve_prompt
+        # Each turn streams its own usage; see record_stream_usage.
+        self.stream_usage_index = nil
+
         api_parameters = api_request_build(prepare_prompt_request, prompt_request_type)
         api_response = instrument("prompt.provider.active_agent") do |payload|
           raw_response = with_exception_handling { api_prompt_execute(api_parameters) }
@@ -193,6 +202,12 @@ def resolve_prompt
           raw_response
         end
 
+        # A stream that deferred its completion has already run
+        # process_prompt_finished (including any tool-call recursion) from
+        # stream_finished!, and holds the response it produced. Calling it a
+        # second time here would re-run that work.
+        return stream_completion_result if stream_completion_result
+
         process_prompt_finished(api_response)
       end
 
@@ -231,7 +246,11 @@ def prepare_prompt_request
       # @return [Hash] API request parameters
       def api_request_build(request, request_type)
         parameters          = request_type.serialize(request)
-        parameters[:stream] = process_stream if request.try(:stream)
+
+        if request.try(:stream)
+          parameters[:stream] = process_stream
+          parameters.deep_merge!(api_stream_usage_parameters)
+        end
 
         if options.extra_headers.present?
           parameters[:request_options] = { extra_headers: options.extra_headers }.deep_merge(parameters[:request_options] || {})
@@ -240,6 +259,20 @@ def api_request_build(request, request_type)
         parameters
       end
 
+      # Extra request parameters needed to make the provider report token usage
+      # while streaming.
+      #
+      # A streaming request returns its usage on a final chunk rather than in a
+      # response body, and several providers only send that chunk when the
+      # request asks for it. Providers that need such a flag override this;
+      # the default asks for nothing, so a provider that reports usage
+      # unconditionally — or not at all — is unaffected.
+      #
+      # @return [Hash]
+      def api_stream_usage_parameters
+        {}
+      end
+
       # @return [Proc] for each response chunk
       def process_stream
         proc do |api_response_chunk|
@@ -258,10 +291,27 @@ def api_prompt_execute(parameters)
           api_prompt_executer.create(**parameters)
         else
           api_prompt_executer.stream(**parameters.except(:stream)).each(&parameters[:stream])
+          stream_finished!
           nil
         end
       end
 
+      # Runs the deferred end-of-generation work once a stream has drained.
+      #
+      # A chunk handler that completed the message marks the generation
+      # pending rather than finishing inline, because providers emit their
+      # usage chunk after the content is done — finishing inline would build
+      # the response before usage is recorded. Handlers that finish inline
+      # leave the flag unset and this is a no-op.
+      #
+      # @return [Object, nil] result of process_prompt_finished, if deferred
+      def stream_finished!
+        return unless stream_completion_pending
+
+        self.stream_completion_pending = false
+        self.stream_completion_result  = process_prompt_finished
+      end
+
       # Returns provider-specific API executer for prompt requests.
       #
       # Since all currently implemented providers use stainless gems, subclasses
@@ -305,6 +355,52 @@ def process_stream_chunk(api_response_chunk)
         fail NotImplementedError, "Subclass expected to implement"
       end
 
+      # Records token usage carried on a streaming chunk.
+      #
+      # The non-streaming path pushes onto usage_stack in resolve_prompt, from
+      # the response body. A streaming request has no response body — the
+      # provider streams chunks and api_prompt_execute returns nil — so usage
+      # would otherwise be lost, and every streamed generation reports zero
+      # tokens and zero cost. Chunk handlers call this when they see usage so
+      # the two paths converge on the same usage_stack.
+      #
+      # Ignores blank and all-zero payloads: providers send `usage: null` on
+      # ordinary content chunks, and pushing those would add empty entries to
+      # a stack that is summed with reduce(:+).
+      #
+      # A streamed usage payload is a running total for the turn, not a delta,
+      # so the turn keeps a single entry that later payloads replace. Chat
+      # Completions sends exactly one, on a final chunk, but Gemini's
+      # OpenAI-compatible endpoint repeats a cumulative usage on every chunk —
+      # summing those would report a turn's tokens many times over. Tool
+      # calling still accumulates across turns: each turn enters resolve_prompt
+      # and starts a fresh entry.
+      #
+      # @param raw_usage [Hash, Object, nil] provider-shaped usage payload
+      # @return [void]
+      def record_stream_usage(raw_usage)
+        return if raw_usage.blank?
+
+        # from_provider_usage only reads hashes, and the stainless gems hand
+        # back model objects (e.g. OpenAI::Models::CompletionUsage), so an
+        # unconverted object is silently dropped.
+        raw_usage = raw_usage.deep_to_h if raw_usage.respond_to?(:deep_to_h)
+        raw_usage = raw_usage.to_h if !raw_usage.is_a?(Hash) && raw_usage.respond_to?(:to_h)
+
+        usage = Common::Usage.from_provider_usage(raw_usage)
+        return if usage.blank?
+        return if usage.total_tokens.to_i.zero? &&
+                  usage.input_tokens.to_i.zero? &&
+                  usage.output_tokens.to_i.zero?
+
+        if stream_usage_index
+          usage_stack[stream_usage_index] = usage
+        else
+          self.stream_usage_index = usage_stack.length
+          usage_stack.push(usage)
+        end
+      end
+
       # Broadcasts stream open event.
       #
       # Fires once per request cycle, even during multi-turn tool calling.
diff --git a/lib/active_agent/providers/open_ai/chat_provider.rb b/lib/active_agent/providers/open_ai/chat_provider.rb
index b1485810..433c3152 100644
--- a/lib/active_agent/providers/open_ai/chat_provider.rb
+++ b/lib/active_agent/providers/open_ai/chat_provider.rb
@@ -26,6 +26,15 @@ def self.prompt_request_type
 
         protected
 
+        # Chat Completions reports a streamed request's token usage on a final
+        # chunk, and only when the request opts in.
+        #
+        # @return [Hash]
+        # @see Base#api_stream_usage_parameters
+        def api_stream_usage_parameters
+          { stream_options: { include_usage: true } }
+        end
+
         # @return [OpenAI::Client::Completions] the API client for chat completions
         # @see Base#api_prompt_executer
         def api_prompt_executer
@@ -100,7 +109,12 @@ def process_stream_chunk(api_response_event)
           # Called Multiple Times: [Chunk<T>, T]<Content, ToolsCall>
           case api_response_event.type
           when :chunk
+            # The usage chunk carries usage and an empty choices array, so it
+            # must be read before choices.first is dereferenced below.
+            record_stream_usage(api_response_event.chunk.try(:usage))
+
             api_message = api_response_event.chunk.choices.first
+            return if api_message.nil?
 
             # If we have a delta, we need to update a message in the stack
             message = find_or_create_message(api_message.index)
@@ -118,8 +132,13 @@ def process_stream_chunk(api_response_event)
             # Returns the full content when complete
             # => {type: :"content.done", content: "Hi there! How can I help you today?", parsed: nil}
 
-            # Once we are finished, close out and run tooling callbacks (Recursive)
-            process_prompt_finished
+            # Close out and run tooling callbacks (Recursive) — but not here.
+            # The usage chunk is emitted *after* content.done, and finishing
+            # at this point builds the response from a usage_stack that has
+            # not received it yet, so a streamed generation reports zero
+            # tokens. Deferred to stream_finished!, which the base provider
+            # calls once the stream has drained.
+            self.stream_completion_pending = true
           when :"tool_calls.function.arguments.delta"
             # => {type: :"tool_calls.function.arguments.delta", name: "get_current_weather", index: 0, arguments: "", parsed: nil, arguments_delta: ""}
           when :"tool_calls.function.arguments.done"
diff --git a/test/fixtures/vcr_cassettes/integration/ollama/chat/native_format_test/test_agent_functions_with_streaming.yml b/test/fixtures/vcr_cassettes/integration/ollama/chat/native_format_test/test_agent_functions_with_streaming.yml
index 666786d5..b21a3695 100644
--- a/test/fixtures/vcr_cassettes/integration/ollama/chat/native_format_test/test_agent_functions_with_streaming.yml
+++ b/test/fixtures/vcr_cassettes/integration/ollama/chat/native_format_test/test_agent_functions_with_streaming.yml
@@ -8,7 +8,7 @@ http_interactions:
       string: '{"messages":[{"role":"user","content":"What is the weather like in
         Boston today?"}],"model":"qwen3","tool_choice":"auto","tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get
         the current weather in a given location","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The
-        city and state, e.g. San Francisco, CA"},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}}],"stream":true}'
+        city and state, e.g. San Francisco, CA"},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}}],"stream_options":{"include_usage":true},"stream":true}'
     headers:
       Accept-Encoding:
       - gzip;q=1.0,deflate;q=0.6,identity;q=0.3
@@ -1765,7 +1765,7 @@ http_interactions:
         MA\"}","name":"get_current_weather"},"type":"function"}]},{"role":"tool","content":"{\"location\":\"Boston,
         MA\",\"unit\":\"fahrenheit\",\"temperature\":\"22\"}","tool_call_id":"call_e26xiwn7"}],"model":"qwen3","tool_choice":"auto","tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get
         the current weather in a given location","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The
-        city and state, e.g. San Francisco, CA"},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}}],"stream":true}'
+        city and state, e.g. San Francisco, CA"},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}}],"stream_options":{"include_usage":true},"stream":true}'
     headers:
       Accept-Encoding:
       - gzip;q=1.0,deflate;q=0.6,identity;q=0.3
diff --git a/test/fixtures/vcr_cassettes/integration/ollama/chat/native_format_test/test_agent_streaming.yml b/test/fixtures/vcr_cassettes/integration/ollama/chat/native_format_test/test_agent_streaming.yml
index 8d0c0040..c0703bbe 100644
--- a/test/fixtures/vcr_cassettes/integration/ollama/chat/native_format_test/test_agent_streaming.yml
+++ b/test/fixtures/vcr_cassettes/integration/ollama/chat/native_format_test/test_agent_streaming.yml
@@ -5,7 +5,7 @@ http_interactions:
     uri: http://127.0.0.1:11434/v1/chat/completions
     body:
       encoding: UTF-8
-      string: '{"messages":[{"role":"developer","content":"You are a helpful assistant."},{"role":"user","content":"Hello!"}],"model":"deepseek-r1:latest","stream":true}'
+      string: '{"messages":[{"role":"developer","content":"You are a helpful assistant."},{"role":"user","content":"Hello!"}],"model":"deepseek-r1:latest","stream_options":{"include_usage":true},"stream":true}'
     headers:
       Accept-Encoding:
       - gzip;q=1.0,deflate;q=0.6,identity;q=0.3
diff --git a/test/fixtures/vcr_cassettes/integration/open_ai/chat/native_format_test/test_agent_functions_with_streaming.yml b/test/fixtures/vcr_cassettes/integration/open_ai/chat/native_format_test/test_agent_functions_with_streaming.yml
index 798f087e..48d3f92a 100644
--- a/test/fixtures/vcr_cassettes/integration/open_ai/chat/native_format_test/test_agent_functions_with_streaming.yml
+++ b/test/fixtures/vcr_cassettes/integration/open_ai/chat/native_format_test/test_agent_functions_with_streaming.yml
@@ -8,7 +8,7 @@ http_interactions:
       string: '{"messages":[{"role":"user","content":"What is the weather like in
         Boston today?"}],"model":"gpt-4.1","tool_choice":"auto","tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get
         the current weather in a given location","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The
-        city and state, e.g. San Francisco, CA"},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}}],"stream":true}'
+        city and state, e.g. San Francisco, CA"},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}}],"stream_options":{"include_usage":true},"stream":true}'
     headers:
       Accept-Encoding:
       - gzip;q=1.0,deflate;q=0.6,identity;q=0.3
@@ -135,7 +135,7 @@ http_interactions:
         MA\"}","name":"get_current_weather"},"type":"function"}]},{"role":"tool","content":"{\"location\":\"Boston,
         MA\",\"unit\":\"fahrenheit\",\"temperature\":\"22\"}","tool_call_id":"call_jbujoPVdWfu8tZ7UjHPotN9G"}],"model":"gpt-4.1","tool_choice":"auto","tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get
         the current weather in a given location","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The
-        city and state, e.g. San Francisco, CA"},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}}],"stream":true}'
+        city and state, e.g. San Francisco, CA"},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}}],"stream_options":{"include_usage":true},"stream":true}'
     headers:
       Accept-Encoding:
       - gzip;q=1.0,deflate;q=0.6,identity;q=0.3
diff --git a/test/fixtures/vcr_cassettes/integration/open_ai/chat/native_format_test/test_agent_streaming.yml b/test/fixtures/vcr_cassettes/integration/open_ai/chat/native_format_test/test_agent_streaming.yml
index 085e6dda..44930da5 100644
--- a/test/fixtures/vcr_cassettes/integration/open_ai/chat/native_format_test/test_agent_streaming.yml
+++ b/test/fixtures/vcr_cassettes/integration/open_ai/chat/native_format_test/test_agent_streaming.yml
@@ -5,7 +5,7 @@ http_interactions:
     uri: https://api.openai.com/v1/chat/completions
     body:
       encoding: UTF-8
-      string: '{"messages":[{"role":"developer","content":"You are a helpful assistant."},{"role":"user","content":"Hello!"}],"model":"gpt-5","stream":true}'
+      string: '{"messages":[{"role":"developer","content":"You are a helpful assistant."},{"role":"user","content":"Hello!"}],"model":"gpt-5","stream_options":{"include_usage":true},"stream":true}'
     headers:
       Accept-Encoding:
       - gzip;q=1.0,deflate;q=0.6,identity;q=0.3
diff --git a/test/fixtures/vcr_cassettes/integration/open_router/native_format_test/test_agent_functions_with_streaming.yml b/test/fixtures/vcr_cassettes/integration/open_router/native_format_test/test_agent_functions_with_streaming.yml
index 23acccb6..70351342 100644
--- a/test/fixtures/vcr_cassettes/integration/open_router/native_format_test/test_agent_functions_with_streaming.yml
+++ b/test/fixtures/vcr_cassettes/integration/open_router/native_format_test/test_agent_functions_with_streaming.yml
@@ -8,7 +8,7 @@ http_interactions:
       string: '{"messages":[{"role":"user","content":"What is the weather like in
         Boston today?"}],"model":"google/gemini-2.0-flash-001","tool_choice":"auto","tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get
         the current weather in a given location","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The
-        city and state, e.g. San Francisco, CA"},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}}],"stream":true}'
+        city and state, e.g. San Francisco, CA"},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}}],"stream_options":{"include_usage":true},"stream":true}'
     headers:
       Accept-Encoding:
       - gzip;q=1.0,deflate;q=0.6,identity;q=0.3
@@ -96,7 +96,7 @@ http_interactions:
         MA\"}","name":"get_current_weather"},"type":"function"}]},{"role":"tool","content":"{\"location\":\"Boston,
         MA\",\"unit\":\"fahrenheit\",\"temperature\":\"22\"}","tool_call_id":"tool_0_get_current_weather_PWzyW2sI3uKsGniIxpGn"}],"model":"google/gemini-2.0-flash-001","tool_choice":"auto","tools":[{"type":"function","function":{"name":"get_current_weather","description":"Get
         the current weather in a given location","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The
-        city and state, e.g. San Francisco, CA"},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}}],"stream":true}'
+        city and state, e.g. San Francisco, CA"},"unit":{"type":"string","enum":["celsius","fahrenheit"]}},"required":["location"]}}}],"stream_options":{"include_usage":true},"stream":true}'
     headers:
       Accept-Encoding:
       - gzip;q=1.0,deflate;q=0.6,identity;q=0.3
diff --git a/test/fixtures/vcr_cassettes/integration/open_router/native_format_test/test_agent_streaming.yml b/test/fixtures/vcr_cassettes/integration/open_router/native_format_test/test_agent_streaming.yml
index dc366aeb..ef9899ca 100644
--- a/test/fixtures/vcr_cassettes/integration/open_router/native_format_test/test_agent_streaming.yml
+++ b/test/fixtures/vcr_cassettes/integration/open_router/native_format_test/test_agent_streaming.yml
@@ -5,7 +5,7 @@ http_interactions:
     uri: https://openrouter.ai/api/v1/chat/completions
     body:
       encoding: UTF-8
-      string: '{"messages":[{"role":"developer","content":"You are a helpful assistant."},{"role":"user","content":"Hello!"}],"model":"openrouter/auto","stream":true}'
+      string: '{"messages":[{"role":"developer","content":"You are a helpful assistant."},{"role":"user","content":"Hello!"}],"model":"openrouter/auto","stream_options":{"include_usage":true},"stream":true}'
     headers:
       Accept-Encoding:
       - gzip;q=1.0,deflate;q=0.6,identity;q=0.3
diff --git a/test/integration/ollama/chat/native_format_test.rb b/test/integration/ollama/chat/native_format_test.rb
index 7eeac885..4fb8446d 100644
--- a/test/integration/ollama/chat/native_format_test.rb
+++ b/test/integration/ollama/chat/native_format_test.rb
@@ -79,6 +79,9 @@ def image_input
                 content: "Hello!"
               }
             ],
+            # Chat Completions only reports a streamed request's token usage
+            # when the request opts in.
+            stream_options: { include_usage: true },
             stream: true
           }
           def streaming
@@ -215,7 +218,7 @@ def web_search
           ###############################################################
           # Extended Example
           ###############################################################
-          FUNCTIONS_WITH_STREAMING = FUNCTIONS.merge(stream: true)
+          FUNCTIONS_WITH_STREAMING = FUNCTIONS.merge(stream_options: { include_usage: true }, stream: true)
           def functions_with_streaming
             prompt(
               model: "qwen3",
diff --git a/test/integration/open_ai/chat/native_format_test.rb b/test/integration/open_ai/chat/native_format_test.rb
index d548dbc9..fdcd640a 100644
--- a/test/integration/open_ai/chat/native_format_test.rb
+++ b/test/integration/open_ai/chat/native_format_test.rb
@@ -100,6 +100,9 @@ def image_input
                 content: "Hello!"
               }
             ],
+            # Chat Completions only reports a streamed request's token usage
+            # when the request opts in.
+            stream_options: { include_usage: true },
             stream: true
           }
           def streaming
@@ -238,7 +241,7 @@ def web_search
           ###############################################################
           # Extended Example
           ###############################################################
-          FUNCTIONS_WITH_STREAMING = FUNCTIONS.merge(stream: true)
+          FUNCTIONS_WITH_STREAMING = FUNCTIONS.merge(stream_options: { include_usage: true }, stream: true)
           def functions_with_streaming
             prompt(
               model: "gpt-4.1",
diff --git a/test/integration/open_router/native_format_test.rb b/test/integration/open_router/native_format_test.rb
index 3f0d97a7..3cbf510b 100644
--- a/test/integration/open_router/native_format_test.rb
+++ b/test/integration/open_router/native_format_test.rb
@@ -96,6 +96,9 @@ def image_input
               content: "Hello!"
             }
           ],
+          # Chat Completions only reports a streamed request's token usage
+          # when the request opts in.
+          stream_options: { include_usage: true },
           stream: true
         }
         def streaming
@@ -310,7 +313,7 @@ def structured_output
           )
         end
 
-        FUNCTIONS_WITH_STREAMING = FUNCTIONS.merge(stream: true)
+        FUNCTIONS_WITH_STREAMING = FUNCTIONS.merge(stream_options: { include_usage: true }, stream: true)
         def functions_with_streaming
           prompt(
             model: "google/gemini-2.0-flash-001",
diff --git a/test/providers/base_provider_stream_usage_test.rb b/test/providers/base_provider_stream_usage_test.rb
new file mode 100644
index 00000000..4c606e7a
--- /dev/null
+++ b/test/providers/base_provider_stream_usage_test.rb
@@ -0,0 +1,112 @@
+# frozen_string_literal: true
+
+require "test_helper"
+require_relative "../../lib/active_agent/providers/mock_provider"
+
+# Streamed token usage: a streaming request has no response body to read
+# usage from — the provider emits it on a final chunk, after the content is
+# done. Without recording that chunk and deferring the end-of-generation work
+# until the stream drains, every streamed generation reports zero tokens (and
+# so zero cost, and no context-pressure estimate downstream).
+class BaseProviderStreamUsageTest < ActiveSupport::TestCase
+  # Stands in for a provider whose stream ends with a usage-bearing chunk.
+  class StreamingMockProvider < ActiveAgent::Providers::MockProvider
+    def self.name = "ActiveAgent::Providers::MockProvider"
+
+    class_attribute :usage_payload, default: nil
+
+    # Mimics the real shape: content arrives first and marks the generation
+    # complete, then a final chunk carries the usage.
+    def api_prompt_execute(parameters)
+      return super unless parameters[:stream]
+
+      message_stack.push({ role: "assistant", content: "ok" })
+      self.stream_completion_pending = true
+      record_stream_usage(self.class.usage_payload)
+      stream_finished!
+      nil
+    end
+  end
+
+  def build(usage:)
+    StreamingMockProvider.usage_payload = usage
+    StreamingMockProvider.new(messages: [ { role: "user", content: "go" } ], stream: true)
+  end
+
+  test "records usage delivered on a final streaming chunk" do
+    response = build(usage: { "prompt_tokens" => 14, "completion_tokens" => 4, "total_tokens" => 18 }).prompt
+
+    assert_equal 14, response.usage.input_tokens
+    assert_equal 4, response.usage.output_tokens
+  end
+
+  test "converts a provider model object rather than dropping it" do
+    # The stainless gems hand back model objects, not hashes; from_provider_usage
+    # only reads hashes, so an unconverted object is silently ignored.
+    object_usage = Struct.new(:to_h).new(
+      { "prompt_tokens" => 30, "completion_tokens" => 9, "total_tokens" => 39 }
+    )
+
+    response = build(usage: object_usage).prompt
+
+    assert_equal 30, response.usage.input_tokens
+    assert_equal 9, response.usage.output_tokens
+  end
+
+  test "ignores the empty usage sent on ordinary content chunks" do
+    provider = build(usage: nil)
+    provider.prompt
+
+    assert_empty provider.usage_stack
+
+    provider = build(usage: { "prompt_tokens" => 0, "completion_tokens" => 0, "total_tokens" => 0 })
+    provider.prompt
+
+    assert_empty provider.usage_stack, "an all-zero payload should not enter the stack"
+  end
+
+  test "a repeated cumulative usage replaces the turn's entry instead of stacking" do
+    # Gemini's OpenAI-compatible endpoint repeats a running total on every
+    # chunk rather than sending one usage at the end of the stream.
+    provider = build(usage: nil)
+    [ 6, 12, 18 ].each do |total|
+      provider.send(:record_stream_usage,
+                    { "prompt_tokens" => total - 2, "completion_tokens" => 2, "total_tokens" => total })
+    end
+
+    assert_equal 1, provider.usage_stack.size
+    assert_equal 16, provider.usage_stack.first.input_tokens, "the last running total wins"
+  end
+
+  test "each turn of a generation records its own usage" do
+    # Tool calling re-enters resolve_prompt per turn, and each turn streams a
+    # usage of its own — those accumulate rather than replacing one another.
+    provider = build(usage: { "prompt_tokens" => 5, "completion_tokens" => 1, "total_tokens" => 6 })
+
+    provider.prompt
+    response = provider.prompt
+
+    assert_equal 2, provider.usage_stack.size
+    assert_equal 10, response.usage.input_tokens
+  end
+
+  # Chat Completions' own opt-in is asserted against a real serialized request
+  # in Providers::OpenAI::Chat::StreamingUsageTest.
+  test "the default provider asks for no extra streaming parameters" do
+    assert_empty build(usage: nil).send(:api_stream_usage_parameters)
+  end
+
+  test "a deferred completion runs process_prompt_finished exactly once" do
+    provider = build(usage: { "prompt_tokens" => 5, "completion_tokens" => 1, "total_tokens" => 6 })
+
+    calls = 0
+    provider.define_singleton_method(:process_prompt_finished) do |*args|
+      calls += 1
+      super(*args)
+    end
+
+    provider.prompt
+
+    assert_equal 1, calls, "resolve_prompt must not re-finish a stream that already completed"
+  end
+end
diff --git a/test/providers/gemini/streaming_lifecycle_test.rb b/test/providers/gemini/streaming_lifecycle_test.rb
index 374a47ca..c7ebdf3c 100644
--- a/test/providers/gemini/streaming_lifecycle_test.rb
+++ b/test/providers/gemini/streaming_lifecycle_test.rb
@@ -132,8 +132,10 @@ def deep_to_h
         )
 
         # Stub process_prompt_finished to just call broadcast_stream_close
+        # The generation finishes when the stream drains, not on content.done
         @provider.stub(:process_prompt_finished, ->(*_) { @provider.send(:broadcast_stream_close) }) do
           @provider.send(:process_stream_chunk, done_event)
+          @provider.send(:stream_finished!)
         end
 
         event_types = @stream_events.map { |e| e[:type] }
@@ -183,8 +185,10 @@ def deep_to_h
         )
 
         # Stub process_prompt_finished to just call broadcast_stream_close
+        # The generation finishes when the stream drains, not on content.done
         @provider.stub(:process_prompt_finished, ->(*_) { @provider.send(:broadcast_stream_close) }) do
           @provider.send(:process_stream_chunk, done_event)
+          @provider.send(:stream_finished!)
         end
 
         refute @provider.send(:streaming), "streaming should be false after close"
diff --git a/test/providers/ollama/streaming_lifecycle_test.rb b/test/providers/ollama/streaming_lifecycle_test.rb
index 4db2300b..3ffc6b22 100644
--- a/test/providers/ollama/streaming_lifecycle_test.rb
+++ b/test/providers/ollama/streaming_lifecycle_test.rb
@@ -93,8 +93,10 @@ def deep_to_h
         )
 
         # Stub process_prompt_finished to just call broadcast_stream_close
+        # The generation finishes when the stream drains, not on content.done
         @provider.stub(:process_prompt_finished, ->(*_) { @provider.send(:broadcast_stream_close) }) do
           @provider.send(:process_stream_chunk, done_event)
+          @provider.send(:stream_finished!)
         end
 
         event_types = @stream_events.map { |e| e[:type] }
@@ -140,8 +142,10 @@ def deep_to_h
         )
 
         # Stub process_prompt_finished to just call broadcast_stream_close
+        # The generation finishes when the stream drains, not on content.done
         @provider.stub(:process_prompt_finished, ->(*_) { @provider.send(:broadcast_stream_close) }) do
           @provider.send(:process_stream_chunk, done_event)
+          @provider.send(:stream_finished!)
         end
 
         refute @provider.send(:streaming), "streaming should be false after close"
diff --git a/test/providers/open_ai/chat/streaming_lifecycle_test.rb b/test/providers/open_ai/chat/streaming_lifecycle_test.rb
index 3b656fbe..95119034 100644
--- a/test/providers/open_ai/chat/streaming_lifecycle_test.rb
+++ b/test/providers/open_ai/chat/streaming_lifecycle_test.rb
@@ -82,7 +82,7 @@ def deep_to_h
           assert_equal 1, open_events.size, "Expected only one :open event even after multiple chunks"
         end
 
-        test "content.done event triggers :close via process_prompt_finished" do
+        test "a drained stream, not content.done, triggers :close via process_prompt_finished" do
           # First send a chunk to trigger :open
           chunk = MockChunk.new(
             choices: [ MockChoice.new(index: 0, delta: MockDelta.new(content: "Hi", role: "assistant")) ]
@@ -90,7 +90,7 @@ def deep_to_h
           chunk_event = MockChunkEvent.new(type: :chunk, chunk: chunk)
           @provider.send(:process_stream_chunk, chunk_event)
 
-          # Then send content.done event which triggers process_prompt_finished
+          # Then send content.done event, which only marks the generation complete
           done_event = MockContentDoneEvent.new(
             type: :"content.done",
             content: "Hi there!",
@@ -101,6 +101,11 @@ def deep_to_h
           # This avoids the nil request issue while testing the streaming lifecycle
           @provider.stub(:process_prompt_finished, ->(*_) { @provider.send(:broadcast_stream_close) }) do
             @provider.send(:process_stream_chunk, done_event)
+
+            assert_empty @stream_events.select { |e| e[:type] == :close },
+                         "content.done must not finish the generation: the usage chunk arrives after it"
+
+            @provider.send(:stream_finished!)
           end
 
           close_events = @stream_events.select { |e| e[:type] == :close }
@@ -133,8 +138,10 @@ def deep_to_h
           )
 
           # Stub process_prompt_finished to just call broadcast_stream_close
+          # The generation finishes when the stream drains, not on content.done
           @provider.stub(:process_prompt_finished, ->(*_) { @provider.send(:broadcast_stream_close) }) do
             @provider.send(:process_stream_chunk, done_event)
+            @provider.send(:stream_finished!)
           end
 
           event_types = @stream_events.map { |e| e[:type] }
@@ -183,8 +190,10 @@ def deep_to_h
           )
 
           # Stub process_prompt_finished to just call broadcast_stream_close
+          # The generation finishes when the stream drains, not on content.done
           @provider.stub(:process_prompt_finished, ->(*_) { @provider.send(:broadcast_stream_close) }) do
             @provider.send(:process_stream_chunk, done_event)
+            @provider.send(:stream_finished!)
           end
 
           refute @provider.send(:streaming), "streaming should be false after close"
diff --git a/test/providers/open_ai/chat/streaming_usage_test.rb b/test/providers/open_ai/chat/streaming_usage_test.rb
new file mode 100644
index 00000000..14410825
--- /dev/null
+++ b/test/providers/open_ai/chat/streaming_usage_test.rb
@@ -0,0 +1,117 @@
+# frozen_string_literal: true
+
+require "test_helper"
+require_relative "../../../../lib/active_agent/providers/open_ai/chat_provider"
+
+module Providers
+  module OpenAI
+    module Chat
+      # End-to-end proof that a streamed generation reports its tokens.
+      #
+      # Chat Completions sends usage on a final chunk that carries an empty
+      # choices array, after the content is done, and only when the request
+      # asked for it. Recording it has to survive the whole chain — the gem's
+      # stream helper, the chunk handler, and the completion deferred until
+      # the stream drains — so this drives real response bodies rather than
+      # hand-built chunk events.
+      class StreamingUsageTest < ActiveSupport::TestCase
+        include WebMock::API
+
+        ENDPOINT = "https://api.openai.com/v1/chat/completions"
+
+        USAGE = { prompt_tokens: 14, completion_tokens: 9, total_tokens: 23 }
+
+        # The order a real Chat Completions stream arrives in: content, then
+        # the finish_reason chunk, then usage on a chunk with no choices.
+        CHUNKS = [
+          { choices: [ { index: 0, delta: { role: "assistant", content: "" }, finish_reason: nil } ] },
+          { choices: [ { index: 0, delta: { content: "Hi" }, finish_reason: nil } ] },
+          { choices: [ { index: 0, delta: { content: " there!" }, finish_reason: nil } ] },
+          { choices: [ { index: 0, delta: {}, finish_reason: "stop" } ] },
+          { choices: [], usage: USAGE }
+        ]
+
+        # Gemini's OpenAI-compatible endpoint ignores the spec and repeats a
+        # running usage total on every chunk instead of sending one at the end.
+        CUMULATIVE_CHUNKS = CHUNKS[0..3].each_with_index.map { |chunk, index|
+          chunk.merge(usage: USAGE.transform_values { |count| ((count * (index + 1)) / 4.0).ceil })
+        } + [ CHUNKS.last ]
+
+        def sse_body(chunks)
+          chunks.map { |chunk|
+            envelope = {
+              id:      "chatcmpl-stream-usage",
+              object:  "chat.completion.chunk",
+              created: 1_761_502_994,
+              model:   "gpt-4o-mini"
+            }
+
+            "data: #{envelope.merge(chunk).to_json}\n\n"
+          }.join + "data: [DONE]\n\n"
+        end
+
+        def stub_stream(chunks)
+          stub_request(:post, ENDPOINT).to_return(
+            status:  200,
+            headers: { "Content-Type" => "text/event-stream" },
+            body:    sse_body(chunks)
+          )
+        end
+
+        setup do
+          @stream_events = []
+
+          stub_stream(CHUNKS)
+        end
+
+        def stream_prompt
+          ActiveAgent::Providers::OpenAI::ChatProvider.new(
+            service:  "OpenAI",
+            api_key:  "test-api-key",
+            model:    "gpt-4o-mini",
+            messages: [ { role: "user", content: "Hello" } ],
+            stream:   true,
+            stream_broadcaster: ->(_message, _delta, event_type) { @stream_events << event_type }
+          ).prompt
+        end
+
+        test "a streamed generation reports the usage from its final chunk" do
+          response = stream_prompt
+
+          assert_equal 14, response.usage.input_tokens
+          assert_equal 9,  response.usage.output_tokens
+          assert_equal 23, response.usage.total_tokens
+        end
+
+        test "usage repeated on every chunk counts once, not once per chunk" do
+          stub_stream(CUMULATIVE_CHUNKS)
+
+          response = stream_prompt
+
+          assert_equal 14, response.usage.input_tokens
+          assert_equal 9,  response.usage.output_tokens
+          assert_equal 23, response.usage.total_tokens
+        end
+
+        test "the streamed content still arrives alongside the usage" do
+          assert_equal "Hi there!", stream_prompt.message.content
+        end
+
+        test "the deferred completion still closes the stream exactly once" do
+          stream_prompt
+
+          assert_equal 1, @stream_events.count(:close)
+          assert_equal :close, @stream_events.last, "the close must come after the last update"
+        end
+
+        test "the request opts in to usage reporting" do
+          stream_prompt
+
+          assert_requested :post, ENDPOINT, times: 1 do |request|
+            JSON.parse(request.body).dig("stream_options", "include_usage") == true
+          end
+        end
+      end
+    end
+  end
+end